Random dynamical systems with microstructure Renato Feres Washington University, St. Louis ESI, July 2011 1/37
Acknowledgements Different aspects of this work are joint with: Tim Chumley (Central limit theorems) Scott Cook (Random billiards with Maxwellian limits) Jasmine Ng (Spectral properties of Markov operators) Gregory Yablonsky (Earlier engineering work) Hong-Kun Zhang (Current work on most of these topics) 2/37
Plan of talk Billiards with random microstructure Billiards: Hamiltonian flows on manifolds with boundary Microstructure: geometric structure on the boundary The Markov operator Derived from the random microstructure Defines generalized billiard reflection Properties of the Markov operator Stationary distributions Spectral gap and moments of scattering The billiard Laplacian Conditioning Case studies CLT and Diffusion 3/37
Random billiards with microstructure 4/37
Random billiards with microstructure 5/37
Wall and molecule subsystems Configuration spaces: Riemannian manifolds with corners and potential functions: U wall M wall Ê U Mol M mol Ê M wall, M mol = M mol Ê Ì k The total system has configuration space M and potential U M Ê. When subsystems sufficiently far away, M M wall M mol and U= U wall + U mol. Outside of product region = interaction zone. Motion: c (t) dt = grad c(t) U with specular collisions at boundary. 6/37
Interaction region (microscopic): definitions S = M mol {0} Ì k M wall boundary of inter. zone; E(q,v) = 1 2 v 2 q + U(q) energy function on N = TM; N S = T S M, N(E) = E 1 (E), N S (E) = N S N(E); θ v (ξ) = v,dτ v ξ contact form on N; dθ symplectic form on N; X E Hamiltonian vector field: X E dθ= de; η =(grad E)/ grad E 2 (in Sasaki metric on N); Ω =(dθ) m Liouville volume form on N; Ω E = η Ω flow invariant volume on energy surfaces; T N S N S the return (billiard) map to S. 7/37
Geodesic flow example (M wall trivial) M wall = single point M mol ={0,1} M mol ={0,1} Ê Ì 1 S={0,1} {0} Ì 1 Potentials are constant 8/37
Example: a dumbbell molecule M wall = single point M mol = SO(2) M mol = SO(2) Ê Ì 1 S= SO(2) {0} Ì 1 Potentials are constant 9/37
Example with potential (M mol trivial) Coordinates: x= m 1 /m(x 1 l/2), y= m 2 /mx 2. E(x,y,ẋ,ẏ)= m 2 (ẋ2 + ẏ 2 + k m 1 x 2 ). 10/37
Random microstructures à la Gromov B = billiard table: Riemannian manifold with boundary; F( B) orthonormal frame bundle with group O(k); N wall = TM wall state space of wall system; V = P(O(k) N wall ) space of probability measures; V is naturally an O(k)-space. Random microstructure on B: O(k)-equivariant map G F( B) P(O(k) N wall ). Example: G =(ξ, ζ) constant, where ξ P(O(k)) is rotation invariant and ζ is Gibbs canonical distribution. 11/37
Gibbs canonical distribution at temperature T An invariant volume form on N wall of physical significance: ζ = e βe Z(β) ΩE wall de where β = 1/κT. (κ = Boltzmann constant.) Density ρ is obtained by maximizing Boltzmann entropy: H(ρ) = Nwall ρlogρω wall under constraint Nwall EρΩ wall = E 0. (β= Lagrange multiplier.) Maximal uncertainty about state given mean value of E. 12/37
The Markov operator P of a microstructure À = half-space in dimension k + 1; N + mol = TM mol À; π N + S = N+ mol Ì k M wall N + mol projection to first factor; λ P(Ì k ) Lebesgue; ζ P(N wall ) a fixed probability (say, the Gibbs measure); T N S + N+ S the return map. Define the map P P(N + mol ) P(N+ mol ), by µ µp =(π T) (µ λ η). 13/37
Markov chains (dynamics under partial state info) Standard finite state Markov chains with detailed balance: M is a groupoid, V is the set of units, T is the inverse operation, v η v are the transition probabilities, and µ η is T-invariant. 14/37
Example of P (constant speed) Transition probabilities operator: (Pf)(θ)= 0 1 f(ψ θ (r)) dr. Surface microstructure defined by a billiard table contour. The coordinate r is random (uniform between 0 and 1). 15/37
Stationary distributions Definition: µ P(N + mol ) is stationary if µp= µ. Theorem Let P P(N + mol) P(N + mol) be the Markov operator associated to the Gibbs canonical distribution on N wall with temperature parameter β. Then the Gibbs canonical distribution on N + mol with the same parameter β is stationary. Proof. Use e β(e mol +E wall ) = e βe mole βe wall and invariance of the symplectic volume form on N S (E) under the return map T. 16/37
Example: Let Ce β 2 m 1w 2 dw ds be fixed state of wall Equilibrium state of molecule: dµ(v)=ccosθ v 2 e β 2 m2 v 2 dθ d v If no moving parts (fixed speed v =1): dµ(θ)= 1 cosθ dθ. 2 No dependence on shapes. 17/37
The operator P on functions Define action of P on functions by ν(p f) =(νp)(f). Definition The molecule-wall system is symmetric if there are volume preserving automorphisms J and K of N + S that: respect the product N + S = N+ mol N wall; induce the same map J on N + mol ; J T= T 1 J (time reversibility) K T= T K (symmetry). Let µ be the stationary measure and H = L 2 (N + mol,µ). Theorem If system is symmetric, P is a self-adjoint operator on H of norm 1. The symmetry condition essentially always holds. Typically, we find in the examples that P is a Hilbert-Schmidt operator. 18/37
Problem: relate structural features and spectrum of P For example, in purely geometric settings (no moving parts on the wall, no potentials) want to relate shape and spectrum. Of special interest: spectral gap. 19/37
Case studies (analytical and numerical) A simple two masses system; Adding a quadratic potential; Billiard systems with no energy exchange; The method of conditioning; Moments of scattering and spectral gap; Systems with weak scattering and the billiard Laplacian. 20/37
A simple two-masses system - I Main system parameter: γ = m 2 m 1 = tanα. 21/37
A simple two-masses system - II Define: x= m 1 m x 1, w = m 1 m v 1, dζ(w) = C exp( 1 2 w2 /σ 2 )dw dx; Theorem P has a unique stationary distribution µ on(0, ), given by dµ(v)=σ 2 v exp( v2 2σ 2)dv. P is a Hilbert-Schmidt operator on L 2 ((0, ),µ) of norm 1; ηp n µ exponentially intv-norm for all initial η. If φ is C 3 on(0, ), then (P γ ϕ)(z) ϕ(z) (Lφ)(z) = lim =( 1 γ 0 2γ 2 z z)ϕ (z)+ ϕ (z) holds for all z> 0. 22/37
A simple two-masses system - III 1 probability density 0.5 0 1 2 3 4 5 speed Evolution of an initial probability measure, µ 0, having a step function density. The graph in dashed line is the limit density v exp( v 2 /2) and the other graphs, from right to left, are the densities of µ 0 P n at steps n=1,10,50,100. Here m 1 /m 2 = 100. 23/37
A simple two-masses system - IV 1 0 0 1 2 3 4 5 speed Comparison of the second eigendensity of P (numerical) and the second eigendensity of the billiard Laplacian L:(1 z 2 /2)ρ(z). Used γ= 0.1; the numerical value for the second eigenvalue of P was found to be 0.9606, to be compared with 1+2γ 2 ( 2)=0.9600 derived from eigenvalue 2 of L. 24/37
A simple two-masses system - V 0.16 0.12 spectral gap 0.08 0.04 0 0 0.04 0.08 0.12 0.16 0.2 mass ratio parameter γ Asymptotics of the spectral gap of P for small values of the mass-ratio parameter γ. The discrete points are the values of the gap obtained numerically. The solid curve is the graph of f(γ)=4γ 2, suggested by comparison with L. 25/37
The bumps family - I (parameter K= l/r>0) dµ(θ)= 1 2 sinθ dθ and H= L2 ([0,π],µ); P K = the Markov operator for bumps with curvature K. Theorem P K is a self-adjoint, compact operator on H of norm 1; For small K, the spectral gap of P K is g(k)= 1 3 K2 + O(K 3 ); 26/37
The billiard Laplacian The (reduced) billiard map is an area-preserving map T S 2 S 2. Regard P K as defined on L 2 (S 2,A). Let be the spherical Laplacian. Theorem Let Φ be a compactly supported smooth function on S 2 {N,S} invariant under rotations about the z-axis in R 3. Then P K Φ Φ= K2 6 Φ+O(K3 ). 27/37
Moments of scattering for bumps family Define the jth moment of scattering E j (θ)=e θ [(Θ θ) j ] where Θ is the random post-collision angle given θ. and PΦ Φ= n j=1 Φ(n) n! E j + O(E n+1 ) if Φ smooth. Proposition If sinθ>3k/2 (middle range of angles), the moments satisfy: If n is odd, E n (θ)= Kn+1 2(n+2) cotθ+o(kn+3 ); If n is even, E n (θ)= Kn n+1 + O(Kn+2 ). It follows that P K Φ Φ = 1 3 K2 1 2sinθ d dθ (sinθdφ dθ )+O(K) 28/37
E 2 (θ)/γ K for K= 2/3,1/3,1/6 E(Θ K θ) 2 /(spectral gap) 1 (constant function) 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.5 1 1.5 2 2.5 3 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.5 1 1.5 2 2.5 3 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0 0.5 1 1.5 2 2.5 3 This gives an asymptotic interpretation of the spectral gap as the mean square deviation from specularity. 29/37
General bumps family, big K Theorem For the general bumps family P is quasi-compact. 0.4 0.3 spectral gap 0.2 0.1 4 10 16 22 Normalized curvature K The solid line is the graph of 2/K and the values marked with an asterisk are numerically obtained. 30/37
The technique of conditioning Wall does not affect J-even eigenvalues, but brings J-odd eigenvalues closer to 0. 0.5 0.45 0.4 spectral gap 0.35 0.3 0.25 0.2 0.1 0.2 0.3 0.4 0.5 height of vertical wall 31/37
More complicated shapes 0.3 spectral gap 0.2 0.1 0.5 0 0.5 distance between bumps, h 32/37
P does not specify shape (up to homothety) Two billiard cells with the same Markov operator: 33/37
Gas transport in channels - 3 levels of description Microscopic model (deterministic motion) Random flight in channel (Markov process on set of directions) Diffusion limit: gas concentration u(x, t) along R should satisfy u t = u D 2 x 2 Relate: (1) microstructure, (2) spectrum of P, (3) diffusion constant. 34/37
Transition to diffusion the Central Limit Theorem Consider the following experiment. Let r= radius of channel; v= constant particle speed; L= half channel length; Release the particle from middle point with distribution ν. Measure the expected exit time, τ(al,r,v) as a. Proposition Suppose P on L 2 ([0,π],µ) has positive spectral gap and µ is ergodic for P. Then where D= 4rv π ξ(p). a 2 τ(al,r,v) 1 D lna We wish to understand how D depends on P. 35/37
D and the spectrum of P Let Π be the (projection-valued) spectral measure of P on[ 1,1]: P= 1 1 Fix β> 1 and let Z a = Zχ { Z a/ln β a}. λ dπ(λ). Spectral measure of Z on[ 1,1]: Π Z ( )=lim a 1 lna Z a,π( )Z a. Theorem D 0 = diffusion const. for i.i.d. process with angle distribution µ. Then D=D 0 1 1 1+λ 1 λ dπ Z(λ). If P has discrete spectrum, Π Z (λ i ) = lim a 1 lna Z a,φ i 2. 36/37
An elementary example θ= initial angle, define integer k, s [0,1), and probability p: 2h btanθ = k+ s, p= s if k is odd 1 s if k is even 1-p h b θ p θ π θ p 1-p Special case: θ=π/4, b>2h. Then k= 0,s=2h/b. For a long channel of diameter 2r and particle speed v, the random flight tends to Brownian motion with σ 2 = 2rv( b 2h 1). An application of the central limit theorem for Markov chains. 37/37